New Hypothesis Links Reward Prediction Errors to Autism Therapy Success

A new hypothesis-and-theory article published in Frontiers in Behavioral Neuroscience proposes that the effectiveness of reinforcement-based therapies for autistic and neurodevelopmental populations may hinge on a single measurable property of the learning signal itself. The author, Sebastian G. D. Cesarano of the Postgraduate School of Integrated Gestalt Psychotherapy in Torre Annunziata, Italy, calls this property Reward Prediction-Error Fidelity, and argues it could serve as a formal bridge between the practical technology of Applied Behavior Analysis (ABA) and the computational neuroscience of reinforcement learning. ABA has long developed effective methods of reinforcement delivery, including Discrete Trial Training, Natural Environment Teaching, and token economies, largely on functional-behavioral grounds without a unified computational account of why these procedures work.

The new hypothesis connects the two traditions with a testable claim: the effectiveness of any reinforcement procedure depends on the quality of the learning signal that its delivery produces on each occasion. Reward Prediction-Error Fidelity is decomposed into four dimensions: temporal credit, contingency, prediction, and outcome value. Each dimension corresponds to a construct that behavior analysts have studied for decades, from the timing of reinforcement to the reliability of the response-reinforcer relation and the subjective value of the reinforcer itself.

The novelty lies not in the individual constructs but in how the hypothesis combines them. The framework is formalized within an actor-critic model equipped with eligibility traces, a standard architecture in computational reinforcement learning in which recently active states and actions remain temporarily eligible for credit when a reward arrives. In this formalization, the four dimensions enter the learning update multiplicatively.

That mathematical choice carries two distinctive implications. First, the dimensions cannot compensate for one another: a reinforcer with excellent timing but poor contingency cannot be rescued by increasing its value, because a multiplicative combination collapses toward zero when any factor is weak. Second, the effect of each dimension depends on the others and on the learner's own history, which distinguishes the hypothesis from steady-state quantitative models of reinforcement that treat parameters as stable across experience.

This combination rule is what the author identifies as the hypothesis's empirical signature. It separates the proposal from the separate behavioral constructs it draws upon, and it generates predictions that differ quantitatively from those of competing models. The article compares the hypothesis with procedural fidelity as traditionally measured in ABA, with quantitative behavioral theory, and with alternative computational accounts including attentional models, timescale-invariant models, successor-representation approaches, and Bayesian learning frameworks.

To make the hypothesis genuinely testable, the article derives nine predictions. For each prediction, the author specifies which result would favor the hypothesis and which result would favor its strongest competitor. This dual-outcome structure is designed to allow the field to falsify or refine the proposal rather than merely accumulate confirming examples.

The research programme is outlined in stages and is explicitly designed to be compatible with single-case experimental designs, the methodological mainstay of behavior-analytic research with small and heterogeneous clinical populations. The application to autistic learners receives particular attention. The article discusses differences in reward processing and in volatility estimation, the degree to which learners track how stable their environment's reward contingencies are, as factors that could modulate Reward Prediction-Error Fidelity in individual children.

It also emphasizes the role of learner assent, positioning the hypothesis within an integrative framework in which the quality of the learning signal is considered alongside the learner's engagement and agreement with the procedure. Because the article is a hypothesis-and-theory contribution rather than an empirical study, its claims are proposals awaiting experimental evaluation, not established findings. No clinical outcomes are reported, and the author notes that all claims are solely those of the author.

The value of the work lies in its precision: by specifying a formal mechanism, a combination rule, and nine discriminating predictions, it converts a long-standing conceptual gap between ABA practice and reinforcement learning theory into a set of concrete experiments that either tradition could conduct.

Subject of Research: A testable hypothesis linking reward prediction-error fidelity in reinforcement learning to Applied Behavior Analysis for autism and neurodevelopmental conditions

Article Title: Reward Prediction-Error Fidelity: a testable hypothesis linking reinforcement learning mechanisms to applied behavior analysis for autism and neurodevelopmental conditions

Article References: Cesarano, S. G. D. (2026). Reward Prediction-Error Fidelity: A Testable Hypothesis Linking Reinforcement Learning Mechanisms to Applied Behavior Analysis for Autism and Neurodevelopmental Conditions. Frontiers in Behavioral Neuroscience, 20, 1974577. https://doi.org/10.3389/fnbeh.2026.1974577

Image Credits: AI Generated

DOI: 10.3389/fnbeh.2026.1974577

Keywords: Applied Behavior Analysis, reinforcement learning, reward prediction error, autism, neurodevelopmental conditions, credit assignment, eligibility traces, actor-critic model, token economies, single-case experimental designs, quantitative behavioral theory, learner assent